from Hacker News

LLMunix: A Markdown OS Experiment Inspired by Karpathy's "LLMs as Computers"

by matiasmolinas on 6/19/25, 12:17 PM with 1 comments

• by matiasmolinas on 6/19/25, 12:17 PM

  Andrej Karpathy's recent AI startup school talk described LLMs as "new computers programmed in English" with properties of operating systems. We built LLMunix to explore these concepts practically.

    Core concept: Everything is defined in markdown documents that Claude interprets as functional system components - no traditional code, just natural language specifications.
  
    Key experiments:
    - Pure markdown OS where .md files become agents, tools, and system processes
    - Sentient state architecture with behavioral constraints that evolve dynamically
    - Memory-driven learning where each execution contributes to behavioral intelligence
    - Agent/tool duality for autonomous workflows
  
    Example: llmunix execute: "Monitor 5 tech sources and generate intelligence briefing" - the system adapts its approach based on API limitations, user sentiment, and historical patterns.
  
    This is early-stage research into how far we can push natural language computing. The results suggest Karpathy's vision of LLM operating systems might be closer than expected.
  
    Links:
    - Karpathy's talk: https://x.com/karpathy/status/1935518272667217925
    - LLMunix repo: https://github.com/EvolvingAgentsLabs/llmunix
  
    Looking for feedback from the HN community on this approach to LLM-native computing.